Target Detection In Clutter Based On Saliency Feature

Along with the rapid development of information technology, the function of image as the main carries of information becomes more and more significant. It contains so much information that it brings difficulties in manual processing. Fortunately, in every tasks of image analysis, it is only small part information that attracts human's interest really. So it is reasonable to import and research selective attention mechanism in human vision system. It can focus finite computing resource on regions of interest to improve the efficiency of image processing, and also suppress the information unrelated to the task effectively to enhance the processing result.It is the primary process to select and extract a series of image features which can weigh the vision saliency of candidate regions accurately. The main goal of this paper is to research how to select and extract saliency features in clutter. Based on the analysis of human vision attention mechanism theory, we research the extraction and description methods of four kinds of saliency features mainly, such as local generalized symmetry, texture and corner, which are self saliency features based on the inner attribute, and then an improved small target detection algorithm based on local symmetry and a target detection approach based on corner voting were proposed, and single feature's localization. Then a general visual computation model based on the difference between inner and exterior attribute was researched, which is based on the Center-Surround operator which extracts relative saliency features, and then its deficiency was analyzed. To the localization of single feature and deficiency of multi-feature integration in the general visual computation model, combined with the serial information processing mechanism in human visual system, we proposed a target description model based on hierarchical saliency features, and a target detection algorithm based on the model. The algorithm imports attention mechanism, and combines both bottom-up and top-down vision attention factors. It extracts several simple saliency features on multi-scale, and uses the features to detect targets singly to get detection probabilities and false detection probabilities, which determine the saliency order of features. And then features are processed hierarchically in target detection under the saliency order. Experiments indicate that the approach proposed is effective and the detection result is consistent with physiological evidence.